When developers build an Android app using generative artificial intelligence tools, the choice of coding agent dictates how often human intervention is required, according to a recent evaluation published on makeuseof.com. In a column for the outlet, developer Bryan M. Wolfe tested three distinct AI systems—Anthropic’s Claude Code, OpenAI’s Codex, and Google’s Antigravity—to program a travel application from scratch without writing any custom code himself.
Building Solo at Sea with Three AI Agents
The test required each AI agent to construct an Android application named Solo at Sea, which serves as a mobile companion for the travel blogging site GoingSolo.Life. The prompt specified Kotlin with Jetpack Compose, bundled sample data instead of live scraping, and a specific navy, gold, rust, and cream color scheme. Each tool received the exact same prompt, word for word, with the objective to build the app, install it over USB, and launch it on a Google Pixel 9 Pro XL running Android 17. The resulting applications were largely interchangeable in appearance and function, but the user experience during the development process varied significantly across platforms.
Claude Code Claims the Speed Record
Running on a Claude Pro subscription powered by Anthropic’s Opus 5.5 model, Claude Code finished the build in just under nine minutes, making it the fastest of the three tested tools. Makeuseof.com reported that Claude Code requested human permission only about three times during the entire build, limiting its queries to basic file and folder access. It was also the only agent that left the final application launch step to the user, requiring the developer to find the newly installed icon in the Android app drawer.
Codex and Antigravity Trade Speed for Oversight
OpenAI’s Codex, accessed via a ChatGPT Plus account using GPT-6 -Astgra, completed the project in just over 14 minutes while prompting the user for permission approximately 10 times. Codex operates inside a sandbox environment, pausing frequently when executing commands outside its immediate boundary, such as downloading build tools or communicating with a device over USB. Meanwhile, Google’s Antigravity, running on a Gemini Pro subscription with the Gemini 3.8 Flash model, took over 22 minutes and required about 30 permission prompts because the reviewer enforced a strict review mode on every terminal command. Both Codex and Antigravity automatically launched the completed application on the test device once the build concluded.
Frequently Asked Questions
- What framework and programming language did the AI agents use to build the app?
- The prompt instructed all three AI tools to use Kotlin combined with Jetpack Compose.
- How many working apps has the author vibe coded in total?
- The author reported vibe coding and re-coding seven working projects over a three-month period.
- Which cruise lines are featured in the Solo at Sea application?
- The application covers three specific lines sourced from the site’s cruise hub: Royal Caribbean, Celebrity, and Viking.
Keep reading